real wages and the business cycle in germanyftp.iza.org/dp5199.pdf · real wages and the business...
TRANSCRIPT
DI
SC
US
SI
ON
P
AP
ER
S
ER
IE
S
Forschungsinstitut zur Zukunft der ArbeitInstitute for the Study of Labor
Real Wages and the Business Cycle in Germany
IZA DP No. 5199
September 2010
Martyna MarczakThomas Beissinger
Real Wages and the
Business Cycle in Germany
Martyna Marczak University of Hohenheim
Thomas Beissinger
University of Hohenheim and IZA
Discussion Paper No. 5199 September 2010
IZA
P.O. Box 7240 53072 Bonn
Germany
Phone: +49-228-3894-0 Fax: +49-228-3894-180
E-mail: [email protected]
Any opinions expressed here are those of the author(s) and not those of IZA. Research published in this series may include views on policy, but the institute itself takes no institutional policy positions. The Institute for the Study of Labor (IZA) in Bonn is a local and virtual international research center and a place of communication between science, politics and business. IZA is an independent nonprofit organization supported by Deutsche Post Foundation. The center is associated with the University of Bonn and offers a stimulating research environment through its international network, workshops and conferences, data service, project support, research visits and doctoral program. IZA engages in (i) original and internationally competitive research in all fields of labor economics, (ii) development of policy concepts, and (iii) dissemination of research results and concepts to the interested public. IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion. Citation of such a paper should account for its provisional character. A revised version may be available directly from the author.
IZA Discussion Paper No. 5199 September 2010
ABSTRACT
Real Wages and the Business Cycle in Germany* This paper establishes stylized facts about the cyclicality of real consumer wages and real producer wages in Germany. As detrending methods we apply the deterministic trend model, the Beveridge-Nelson decomposition, the Hodrick-Prescott filter, the Baxter-King filter and the structural time series model. The detrended data are analyzed both in the time domain and in the frequency domain. The great advantage of an analysis in the frequency domain is that it allows to assess the relative importance of particular frequencies for the behavior of real wages. In the time domain we find that both real wages display a procyclical pattern and lag behind the business cycle. In the frequency domain the consumer real wage lags behind the business cycle and shows an anticyclical behavior for shorter time periods, whereas for longer time spans a procyclical behavior can be observed. However, for the producer real wage the results in the frequency domain remain inconclusive. JEL Classification: E32, C22, C32, J30 Keywords: real wages, business cycle, frequency domain, time domain, Germany,
trend-cycle decomposition, structural time series model, phase angle Corresponding author: Martyna Marczak University of Hohenheim Department of Economics Schloss, Museumsfluegel D-70593 Stuttgart Germany E-mail: [email protected]
* We thank Anita C. Bott, Jacques J.F. Commandeur, Irma Hindrayanto, Kurt Jetter, the participants of the WIEM 2010 conference in Warsaw and the participants of the annual congress of the German Economic Association 2010 in Kiel for helpful comments.
1 Introduction
At least since Keynes (1936) claimed in his General Theory that an increase in em-
ployment can only occur with a simultaneous decline in real wages, macroeconomists
are debating about whether real wages are anticyclical, procyclical or do not exhibit
any systematic relationship with the business cycle. A clarification of this issue
could shed some light on the main sources of macroeconomic shocks and thereby
be of some use in judgements about the empirical relevance of conflicting macroe-
conomic theories. A clearer empirical picture about the adjustment of real wages
over the business cycle also helps in identifying the sources and features of wages
and labor cost dynamics and therefore is of great relevance for monetary policy.
This paper contributes to the literature on the adjustment of aggregate real
wages over the business cycle in several ways. First, we analyze the comovements
between real wages and the cycle not only in the time domain, but also in the
frequency domain. So far, most studies have focussed on the time domain approach
and described the comovements between variables by traditional cross–correlations
measures. However, as has been pointed out by Hart et al. (2009), in the time
domain the observed cyclical behavior of the real wage hides a range of economic
influences that give rise to cycles of different length and strength, thereby producing
a distorted picture of real wage cyclicality. The great advantage of an analysis in
the frequency domain is that it allows to assess the relative importance of particular
frequencies for the behavior of real wages.
Second, it is analyzed whether the empirical results are robust to the method
used to extract the cycle from the data. More specifically, we apply the deterministic
trend model, the Beveridge–Nelson decomposition, the Hodrick–Prescott filter, the
Baxter–King filter and the structural time series model of Harvey (1989) to the time
series of aggregate real wages and gross domestic product. Since it is well known
from the literature that the results may also be influenced by the price deflator used
1
to compute real wages, we take this into account by considering both producer real
wages and consumer real wages.
Third, we analyze the real wage behavior for the economy as a whole whereas
many studies only consider real wages in the manufacturing sector, as for example
in the recent study of the wage dynamics network of the ECB on real wage behavior
in the OECD, see Messina et al. (2009).1 Because of the much larger shares of the
non-manufacturing sector in total output and employment, empirical results for the
economy as a whole are certainly preferable.
Forth, whereas the question of the cyclicality of real wages in the US has been
analyzed in a host of studies (see the surveys of Abraham and Haltiwanger, 1995,
and Brandolini, 1995), surprisingly little systematic empirical evidence exists for
Germany. This paper tries to fill this gap and provides a detailed picture of the
wage dynamics in an economy in which labor unions (still) affect the majority of
employment contracts.
The remainder of the paper is organized as follows. In Section 2 we describe
our data and then analyze the stochastic properties of the time series. In Section 3
different trend–cycle decompositions are applied to consumer real wages, producer
real wages and real GDP. In Section 4 we analyze the comovements between the
particular real GDP cycle and the corresponding real wage cycles in the time and
frequency domain. Section 5 summarizes and concludes.
2 Data and stochastic properties of the series
We use quarterly data for real GDP, consumer real wages and producer real wages
in Germany from 1970.Q1 to 2009.Q1 (157 observations). All series that served to
generate the project data were seasonally adjusted with the Census-X12-ARIMA
1In Messina et al. (2009) also time domain and frequency domain methods are used.
2
procedure. The data prior to 1991.Q1 refers to West Germany and has been linked
to the data of unified Germany using annual averages for 1991. The data selection
is described in more detail in Appendix A. All generated data are represented in
natural logarithms.
Before we undertake the trend-cycle decompositions we study the stochastic
properties of the data. We test for unit roots in real GDP and both real wage series
applying the augmented Dickey-Fuller (ADF) test and the Phillips-Perron test. In
both tests the alternative hypothesis is based on the assumption that the particular
series follows a trend–stationary process since all series exhibit a clear upward course.
The lag length for the ADF test is determined by the Akaike information criterion
(AIC) and Schwartz information criterion (SIC) and its accuracy is then verified
with the Ljung-Box test and the Breusch-Godfrey test. The results of both unit
root tests show that for all series the unit root hypothesis cannot be rejected using
conventional significance levels. Therefore, the underlying stochastic processes are
not covariance stationary. We then apply both unit root tests to the first differences
of the series. Since the null hypothesis can now be rejected, we conclude that all
series are generated by I(1) processes.
3 Identification of the cyclical component
The general framework for the decomposition of each time series into trend and
cycle is provided by the following model:
yt = ygt + yc
t + εt, t = 1, 2, ..., T (1)
where t is a time index and yt represents the natural logarithm of the series under
consideration, i.e. real GDP, consumer real wages or producer real wages. The series
yt is decomposed into trend ygt , cycle yc
t and (possibly) an irregular component εt.
The latter is only relevant in the structural time series model (STSM), whereas it
3
is neglected in the other decomposition methods applied in this paper, namely the
linear trend model with broken trend (LBT), the Beveridge–Nelson decomposition
(BN), the Hodrick–Prescott filter (HP) and the Baxter–King filter (BK). The latter
methods assume the variance of εt to be zero, thereby attributing any disturbance
left in the data after the removal of the trend to the cyclical component.
As a first decomposition method we consider the linear trend model. To check
whether the time series under consideration is subject to structural breaks we apply
the Quandt-Andrews test.2 For all time series the test clearly rejects the hypothesis
of no structural break. The estimated break point is 2002.Q4 for real GDP and
2003.Q1 for both real wages. Based on this result we estimate the following model:
yt = α0 + α1t + β1St,k + νt,
St,k =
t− k, if t > k
0, if t ≤ k,
(2)
where νt is generated by a covariance stationary process which is uncorrelated with
{yt} and St,k reflects the change in the slope of the trend starting with period k.
The results of the OLS estimation of model (2) are reported in Table 1. According
to these findings real GDP grows with a (quarterly) rate of 0.57% until 2002.Q4.
From 2002.Q4 on its growth slows down by 0.47 percentage points. The growth rate
of the consumer real wage equals 0.35% before and -0.33% after the break point.
The growth rate of the producer real wage exceeds that of the consumer real wage
by 0.17 percentage points over the first period. However, after 2003.Q1 it falls more
steeply than in the case of the consumer real wage. The deviations from the growth
path, i.e. the residuals of the estimated model, represent the cyclical component,
hence yct = νt. The examination of the residuals with correlograms and the Ljung-
2See Andrews (1993). This test overcomes the shortcomings of the commonly used Chow test
in that it does not require any previous knowledge about the occurrence of a possible break point.
We choose standard 15% as “trimming” level for this test.
4
Table 1: Estimation of segmented linear trend models for real
GDP and real wages
regressorGDP consumer wage producer wage
coefficients a)
t0.0057 0.0035 0.0052
(104.21) (59.19) (46.87)
Sk,t
−0.0047 −0.0068 −0.0112
(−10.49) (−13.4) (−11.77)
constant5.544 2.385 2.173
(1278.237) (510.194) (246.321)
a) t-values in parentheses. Number of observations: 157. Break point
k = 132 for real GDP and k = 133 for real wages.
Box test indicate that the obtained cycles of real GDP and real wages follow a
persistent AR(1) process. This is confirmed by the results of the ADF test applied
to each of these cycles. The hypothesis of a unit root cannot be rejected at the 5%
significance level in each case. Since the LBT cycles do not satisfy the stationarity
condition which is needed for the comovement analysis, we exclude the LBT cycles
from further analysis.
As has been shown in Section 2, both real GDP and real wages are difference–
stationary processes. For this case, a suitable decomposition method has been sug-
gested by Beveridge and Nelson (1981). The BN decomposition assumes a I(1)
process for the examined series and regards the trend as a prediction of future val-
ues of the series. The decomposition leads to a trend component which is a random
walk with drift and to a covariance stationary cyclical component which are cor-
related with each other. In this respect, the BN decomposition differs from the
5
LBT model with its strong assumption of zero correlation between trend and cycle.
However, the BN decomposition also bears some problems. For example, it requires
an ARMA specification for the examined series but since distinct ARMA models
can suit the data, different forecasts can result from these models. That in turn
implies different trends and cycles. Furthermore, the a priori assumption about the
trend being a random walk is somewhat controversial. Another problematic issue
concerns the variance of the trend that could even exceed that of the series.
The procedure determining the trend and cycle requires truncation of infinite
sums and is associated with heavy computational burden. In order to reduce these
costs, different resolution methods have been proposed in the literature.3 In this
paper, we take the approach of Newbold (1990) which is based on the ARIMA(p, 1, q)
representation of the series.4 The BN cycle can be described as:
yct =
q∑j=1
[zt(j)− µ] + (1− φ1 − ...− φp)−1
p∑j=1
p∑i=j
φi[zt(q − j + 1)− µ], (3)
where zt(k) denotes the k-periods ahead forecast of z = ∆y made in period t. φj is
the AR coefficient at lag j and µ is the mean of the process {zt}. To isolate the cycle
using (3) we have to identify the best ARMA specification for the first differences
of real GDP and real wages. In doing so, we rely on the information criteria (AIC
and SIC) beginning with an AR(4) model for each of the series in first differences.
Since ARMA modeling technique aims at a parsimonious representation we reduce
the number of AR terms and then optionally add some MA terms and compare all
models with regard to the values of AIC and SIC. The initially considered AR(4)
specification turns out to be the most suitable one. Next, we examine the residuals
3See, for example, Cuddington and Winters (1987), Miller (1988) and the more recent work of
Morley (2002).4According to Wold’s theorem, each covariance stationary process has a MA(∞) representation
which is also consistent with an ARMA(p, q) representation. Therefore, each I(1) process in first
differences has an ARMA(p, q) representation.
6
from this model with the Ljung-Box test and the Breusch-Godfrey test. We find
no evidence for serial correlation of the residuals in the case of real GDP as well as
the real producer wage in first differences, respectively. Hence, for these series we
choose an AR(4) model. As for the first differences of the real consumer wage, the
residual autocorrelation vanishes after including an additional lag, so we finally end
up with an AR(5) specification. Inserting the forecasts based on the selected models
in (3) yields the cyclical components of real GDP and real wages.
As the next trend-cycle decompositions we use linear filters, the HP filter and
the BK filter, which have proven popular in macroeconomic applications.5 A great
advantage of these methods can be seen in the fact that they are able to render the
data stationary. They also avoid modeling of the series in contrast to, e.g. , the BN
decomposition. However, the results of both filters are not without problems if they
are applied to series which are generated by nonstationary processes. It has been
shown in the literature that in this case the HP filter induces spurious cycles.6 This
is due to the fact that the frequency components of the resulting series have business
cycle periodicity even though there are no important transitory fluctuations in the
original data. As regarding the critique of the BK filter application for nonstationary
series, Murray (2003) demonstrates that the first difference of an integrated trend
enters the filtered series. As a result, the spectral properties of the filtered series
depend on the trend in the unfiltered series. Because of the nonstationarity of
the analyzed series the cycles obtained with the HP and the BK filter should be
interpreted with some caution.
Finally, we consider structural time series models, which are defined in terms
of unobserved components that have a direct economic interpretation.7 The initial
5As suggested by Hodrick and Prescott (1980), for the HP filter we use the value 1600 for the
smoothing parameter.6See, for example, Cogley and Nason (1995) and Harvey and Jaeger (1993).7See Harvey (1989, pp. 44–49).
7
specification of the model structure is left to the researcher. Within this framework,
the data decide on the characteristics of the particular component. In contrast to the
ad hoc filtering approaches, such as the HP and the BK filter, structural time series
models rely on the stochastic properties of the data. Moreover, as opposed to ARMA
modeling they do not aim at a parsimonious specification. It is quite probable
that a parsimonious ARMA model identified by means of standard techniques (e.g.
correlograms) does not exhibit properties expected from the examined series. For
instance, it could reject cyclical behavior of a series even though such a behavior
does really exist. Unfortunately, finding a “correct” model specification inevitably
also remains a problem in the case of structural time models.
In this paper, we adopt the model as in eq. (1) and assume that the irregular
component εt is normally, independent and identically distributed with variance σ2ε :
εt ∼ NID(0, σ2ε) (4)
The stochastic trend component ygt can be formulated as follows:8
ygt+1 = yg
t + βt + ηt, ηt ∼ NID(0, σ2η)
∆mβt+1 = (1− L)mβt+1 + ζt, ζt ∼ NID(0, σ2ζ )
(5)
The variable βt is the slope of the trend and the scalar m (m = 1, 2, 3, ...) is the
order of the slope. If the slope follows an I(m) process then the trend is I(m + 1).
In case of m = 1 the trend is called local linear trend. Imposing restrictions on the
variances σ2η and σ2
ζ leads to various trend forms. With m = 1 and σ2η = σ2
ζ = 0 one
obtains a deterministic linear trend. The assumption σ2ζ = 0 together with m = 1
implies that the trend is a random walk, whereas σ2η = 0 along with m = 1 results in
a relatively smooth I(2) trend component. If one needs to model an even smoother
trend component the trend is supposed to be of higher order (m > 1). The cycle yct
8See Koopman et al. (2009, S. 55-56).
8
is defined as:9
yc
t+1
yc∗t+1
= ρ
cos(ω) sin(ω)
− sin(ω) cos(ω)
yc
t
yc∗t
+
χt
χ∗t
,
χt
χ∗t
∼ NID(0, σ2
χI2),
(6)
where yc∗t is an auxiliary variable, ω denotes the frequency (0 ≤ ω ≤ π) and ρ is
the damping factor (0 ≤ ρ ≤ 1). The period p of the cycle is therefore p = 2π/ω.
If ω = 0 or ω = π, the VAR(1) process in (6) collapses into an AR(1) process. The
variance σ2χ is given as σ2
χ = σ2c (1 − ρ2), where σ2
c is the variance of the cycle so
that with ρ = 1 the cycle is reduced to a deterministic and covariance stationary
process. For all three series we start with the general formulation of the model with
no variance restriction (model 1), but we constrain the trend specification to the
local linear trend (m = 1). The whole model is estimated by maximum likelihood
with the Kalman filter. The Kalman smoothing provides the estimates of the trend
and cycle component. The estimated model parameters, called hyperparameters,
that refer to real GDP are reported in Table 2. A high value of σ2η relative to σ2
χ
indicates an erratic trend component and a damped cycle component. Since this
result seems rather implausible, in the next step we apply the restriction σ2η = 0
ensuring bigger deviations of the cycle than in the general model (see Table 2, model
2). The cycle extracted from the restricted model does coincide better with the
German history of booms and recessions. The results of a likelihood ratio (LR) test
also confirm the validity of the variance restriction. For the consumer real wage, the
initial model leads to a deterministic cyclical component so in this case we reject the
general model, too (see Table 3). We enforce the cyclical component to be stochastic
9See Koopman et al. (2009, p. 63), Koopman et al. (2008, p. 23) und Harvey and Streibel
(1998). Clark (1989) suggests to describe the cycle as an AR(2) process. Harvey and Trimbur
(2003) generalize the trigonometric version in (6) to cycles of higher order.
9
Table 2: Estimation of the general and restricted trend-cycle model for
real GDP
modelhyperparametera)
R2D
b)
σ2ε σ2
η σ2ζ σ2
χ ρ ω
1) no restrictions 0,865 74,98 0,215 11,734 0,976 0,234 0,0545
2) σ2η = 0 5,032 – 0,449 67,683 0,929 0,194 0,0469
a) The estimated variances have been multiplied by 106.b) The coefficient of determination R2
D is based on the first differences of the ob-
served series.
Table 3: Estimation of three trend-cycle models for the consumer real wage
modelhyperparametera)
R2D
b)
σ2ε σ2
η σ2ζ σ2
χ ρ ω
1) no restrictions, m = 1 24,187 22,883 1,862 0 1 0,499 0,0964
2) σ2η = 0, m = 1 33,166 – 3,298 1,102 0,966 0,468 0,0697
3) σ2η = 0, m = 2 22,482 – 0,004 26,537 0,957 0,153 0,0498
a) The estimated variances have been multiplied by 106
b) The coefficient of determination R2D is based on the first differences of the observed series
by assuming σ2η = 0. However, the irregular term becomes the most important
component in explaining the consumer real wage variation and the cycle has a high
frequency (see Table 3, model 2). These problems are eliminated if we allow for
a smoother trend by setting m = 2 (see Table 3, model 3). We proceed similarly
with the model identification for the producer real wage, hence we choose the trend-
cycle model with σ2η = 0 and m = 2. The estimation results are summarized
10
in Table 4. Following e.g. Harvey and Koopman (1992) and Commandeur and
Table 4: Estimation of three trend-cycle models for the producer real wage
modelhyperparametera)
R2D
b)
σ2ε σ2
η σ2ζ σ2
χ ρ ω
1) no restrictions, m = 1 14,894 61,212 9,687 0 1 0,542 0,119
2) σ2η = 0, m = 1 30,67 – 12,293 10,776 0,928 0,505 0,0901
3) σ2η = 0, m = 2 10,39 – 0,009 73,623 0,947 0,203 0,0914
a) The estimated variances have been multiplied by 106
b) The coefficient of determination R2D is based on the first differences of the observed series
Koopman (2007), we then check all selected models with the following diagnostic
tests: the Ljung-Box autocorrelation test, the Goldfeld-Quandt heteroscedasticity
test and the Bowman-Shenton normality test. In all cases, we cannot reject the
hypothesis of no autocorrelation at the 5% significance level. The heteroscedasticity
test finds evidence against the homoscedasticity assumption only for the consumer
real wage, whereas the Bowman-Shenton test indicates violation of the normality
assumption for all series. Nevertheless, since the main concern of time series analysis
is the autocorrelation problem we can conclude that these models provide a satisfying
specification of the data generating process.
In Figure 1, we depict the cyclical component for real GDP for the different
detrending methods outlined above. One can easily recognize the periods of booms
and recessions that Germany has experienced since 1970. The first recessions occur
as a result of the first and second oil crisis 1974–75 and 1980–82.10 After a relatively
weak recovery in the second half of the 1980s one can observe a clear boom phase
10We do not interpret the large negative values at the beginning of the sample in the case of the
LBT and HP cycles since they could be caused by problematic behavior of these methods at the
bounds.
11
in the early 1990s that is due to German reunification. The next turning point is
reached in 1993 with the beginning of a recession phase initiated by the restrictive
monetary policy of the Deutsche Bundesbank. In the late 1990s one observes a
recovery that may have been caused by the IT boom followed by a recessionary
phase 2001–2005. Afterwards, the economy expands again. This positive course
ends in 2008 because of the world economic and financial crisis leading to a severe
downturn. The various cyclical components of the consumer real wage and the
Figure 1: Cycles of real GDP
LBT BK
HP
1970 1975 1980 1985 1990 1995 2000 2005 2010
−0.05
0.00
0.05LBT BK
HP
BN STSM
1970 1975 1980 1985 1990 1995 2000 2005 2010
−0.050
−0.025
0.000
0.025
0.050BN STSM
producer real wage are compared in Figure 2 and Figure 3, respectively.
For all series, the LBT cycles exhibit the most striking peaks and troughs. It is
12
Figure 2: Cycles of the consumer real wage
LBT BK
HP
1970 1975 1980 1985 1990 1995 2000 2005 2010
−0.05
0.00
0.05LBT BK
HP
BN STSM
1970 1975 1980 1985 1990 1995 2000 2005 2010
−0.05
0.00
0.05
BN STSM
apparent that the BN cycles of both real wages are shifted relative to the STSM,
HP and BK cycles. Moreover, the STSM cycles of both real wages are almost in line
with the HP cycles. This can be explained by the fact that the structural time series
model with a trend of higher order can be associated with a Butterworth filter.11
11Gomez (2001) shows that the HP filter is a Butterworth filter.
13
Figure 3: Cycles of the producer real wage
LBT BK
HP
1970 1975 1980 1985 1990 1995 2000 2005 2010
−0.10
−0.05
0.00
0.05
0.10LBT BK
HP
BN STSM
1970 1975 1980 1985 1990 1995 2000 2005 2010
−0.1
0.0
0.1
BN STSM
4 Comovements of real GDP and real wages
4.1 Time Domain
The analysis of comovements in the time domain between real wage cycles and real
GDP cycles as a reference for the business cycle is a natural approach to detect the
cyclical behavior of real wages. The measure of comovements we adopt here are the
sample cross–correlations between the cycle of each of the real wage series and the
real GDP cycle. We consider not only the contemporaneous relationship but also
analyze whether real wages react with delay or run ahead of cyclical movements in
14
real GDP. We find it a bit misleading that in the literature wages are often classified
as pro– or countercyclical by solely focusing on the contemporaneous correlation.12
Using quarterly data, one can not seriously expect that the main adjustment of real
wages to cyclical changes in GDP will take place in the same period. Therefore, we
classify the considered real wage as procyclical (countercyclical) if the estimated cor-
relation coefficients are positive (negative) taking into account the lead–lag structure
of the examined series. If the estimated correlation coefficients are close to zero the
particular real wage is defined to be acyclical. If the largest sample cross–correlation
occurs at any lead (lag) relative to the GDP cycle we say that the particular real
wage is lagging (leading) the cycle.
Table 5: Contemporaneous and largest sample cross–correlations between the
real GDP cycle and the particular real wage cycle by various decom-
position methods
correlation of methods
GDP with BN HP BK STSM
consumer real wage0,1169 0,0124 0,1438 −0, 1677∗
0, 4879∗(+6) 0, 4572∗(+6) 0, 6346∗(+5) 0, 4099∗(+11)
producer real wage0,0279 −0, 0423 0,0314 −0, 0362
0, 2718∗(+6) 0, 2381∗(+7) 0, 315∗(+7) 0, 2163∗(+10)
Notes: “ * ” indicates statistical significance at the 5% level
The findings are summarized in Table 5. Each cell contains in the first row the
contemporaneous sample cross–correlation between the cycle of the real wage series
and the corresponding real GDP cycle. The value below is the maximum sample
12This focus of the literature on the contemporaneous correlation has also been criticized by
Messina et al. (2009).
15
cross–correlation at the kth lead or lag of the real wage cycle relative to the real GDP
cycle, where k ∈ {−12,−11, ..., 0, ..., 11, 12}. The number in brackets along with the
“+” or “−” sign specifies at which lead or lag of the real wage cycle this maximum
cross–correlation occurs.13 We first consider the results for the consumer real wage.
Except for the STSM cycle, the estimates of the contemporaneous cross–correlation
are positive but statistically insignificant at the 5% level. The low practical signifi-
cance is most apparent in the case of the HP cycle. Considering the leads of the real
wage cycles, we find that for all cycles except for the STSM cycles, the relationship
with the corresponding real GDP cycles is still positive but now becomes significant.
The sample cross–correlations reach their maximum values at the 6th lead (BN and
HP cycles) or 5th lead (BK cycles). In the case of the STSM cycles, there is first a
significant negative sample cross–correlation until the 3rd lead. From the 6th lead,
it takes high positive values that are statistically significant. We find the greatest
cross–correlation at the 11th lead. Examination of the lags of the real wage cycles
reveals that almost all sample cross–correlations are statistically insignificant in the
case of the HP, BK and BN cycles. The significant ones are small compared to
the significant sample cross–correlations at the leads.14 To sum up, the consumer
real wage displays a procyclical pattern and lags the business cycle. The strongest
reaction to the actual economic situation can be observed between the 5th and the
11th quarter.
The behavior of the producer real wage differs somewhat from that of the con-
sumer real wage. All estimated contemporaneous cross–correlations are statistically
insignificant at the 5% level. Furthermore, although there is a similar cyclical pattern
13For clarity reasons, we do not present detailled figures of the lead–lag structure and instead
describe some results verbally.14In the case of the STSM cycles, the significant negative sample cross–correlations emerge at
the first 3 lags. In contrast, the BN cycles are characterized by positive cross–correlations which,
though, are insignificant.
16
as in the case of the consumer real wage, the sample cross–correlations at the leads
of the real wage are not as high. In Table 5 this is evident from the differences in the
maximum cross–correlations between both wages. The sample cross–correlations at
the lags of the producer real wage, with the exception of some lags in the case of the
BN cycles, are statistically insignificant. The analysis leads to the conclusion that
the producer real wage behaves procyclically and lags the business cycle. The main
reaction to the actual economic situation occurs after 6 (BN cycle) to 10 (STSM
cycle) quarters.
4.2 Frequency domain
The above analysis of the comovements between real wages and the cycle in the
time domain might give the impression that the cyclicality of real wages has been
sufficiently characterized. However, the observed behavior of real wages in the time
domain results from the countervailing or/and reinforcing influences of cycles of
different length. As a consequence, if we want to learn something about the behavior
of real wages over the business cycle, we could be misled by looking at the time
domain results alone. In this section, we resort to some spectral analysis concepts
that enable us to assess the relative importance of cycles of different length and
strength and therefore provide a comprehensive picture on the cyclical behavior of
real wages.
We will first give a short introduction to these concepts. The central one is the
spectral representation of a covariance stationary process, also called the Cramer
representation, as a frequency domain counterpart to the Wold representation of
such a process. According to the spectral representation a time series Yt, which
is a single realization of a zero–mean covariance stationary process yt, is regarded
as comprising various superimposed cosine and sine waves each having different
frequency and amplitude. If such a stochastic process yt is discrete and real–valued,
17
it can be described by:15
yt =
∫ π
0
α(ω) cos(ωt)dω +
∫ π
0
δ(ω) sin(ωt)dω, (7)
where α(ω) and δ(ω) are orthogonal complex–valued stochastic processes with zero
mean and equal variances. The variable ω denotes the (angular) frequency. The
coefficients α(ω) and δ(ω) give rise to the stochastic nature of the process in (7), in
that |α(ω)| and |δ(ω)| are random amplitudes, whereas arg{α(ω)} and arg{δ(ω)} de-
scribe random phases of the particular cosine and sine wave. Each wave contributes
to the explanation of the overall variance (power) of this process. This information
is given by the real–valued function s(ω), the so–called spectral density function or,
in short, spectrum:
s(ω) =1
2π
∞∑j=−∞
γj e−iωj
=1
2π
∞∑j=−∞
γj cos(ωj)
=1
2π
[γ0 + 2
∞∑j=1
γj cos(ωj)],
(8)
where γj is the jth autocovariance of the process and i is the imaginary number.
The area under the graph of s(ω) for ω ∈ [−π, π] describes the total variance of the
process.
Since we are primarily interested in the interactions between time series, we now
turn to the multivariate case and consider two series Ykt and Ylt (k, l = 1, 2, .., n, k 6=l). The frequency by frequency relationship between the underlying processes ykt
and ylt can be measured by the cross–spectrum skl(ω):
skl(ω) =1
2π
∞∑j=−∞
γjkl e−iωj
=1
2π
∞∑j=−∞
γjkl[cos(ωj)− i sin(ωj)],
(9)
15See, for example, DeJong and Dave (2007, pp. 41–42) and Priestley (1981, pp. 251–252).
18
where γjkl is the jth cross–covariance of the two processes defined as
γjkl = E[(ykt − µk)(yl,t−j − µl)] (10)
The cross–spectrum, which is a complex–valued function of ω, can be decomposed
into the real part ckl(ω) called cospectrum and the imaginary part qkl(ω) called
quadrature spectrum. In analogy to the spectrum of an individual process, the
area under the cross–spectrum in the range [−π, π] gives the overall covariance of
the two processes. Additionally, as the quadrature spectrum integrates to zero in
this interval, the area under the cross–spectrum is equal to the area under the
cospectrum. According to this, the cospectrum at frequency ω can be interpreted as
the marginal contribution of the components with frequency ω + dω to the overall
covariance between the processes. The quadrature spectrum at this frequency can
serve as an indicator for the out–of–phase covariance since it measures the portion
of the covariance between two processes shifted relative to one another by π/2 which
is attributable to the waves with this frequency. The information contained in both
the cospectrum and the quadrature spectrum is useful in establishing the lead–lag
relationship between two processes. For this purpose, the inferences from both parts
of the cross–spectrum at any frequency can be combined into one quantity, the so–
called phase angle, denoted by θ(ω):
θ(ω) = arctan
[qkl(ω)
ckl(ω)
](11)
Because of the properties of arctangent, the phase angle θ(ω) is a multivalued func-
tion, but it is common to limit its values to the interval (−π, π). The unique value
in (−π, π) and therewith the sign of the phase angle can, however, only be deter-
mined by the signs of the cospectrum and the quadrature spectrum. If θ(ω) takes
on positive values, we say that the component of ykt with frequency ω leads the
corresponding component of ylt. The opposite case is implied by θ(ω) < 0. Both
components are in phase if θ(ω) equals zero. Based on the values of the phase angle
19
we can also make statements about the correlation between ykt and ylt. If the values
of the phase angle range between [−π/2, π/2], we say that ykt and ylt are positively
correlated (procyclical behavior), whereas the values of θ(ω) in the interval [π/2, π]
or [−π/2,−π] indicate a negative relationship (countercyclical behavior) between
them.
In this paper, we focus on the nonparametric approach to the estimation of
spectra and cross–spectra.16 For that purpose, we have to choose a suitable spectral
window and the truncation point of the window. Since, as pointed out by Jenkins
and Watts (1968, p. 280), the spectral estimates are barely affected by the functional
form of the window, we use the Bartlett window and concentrate on finding an
appropriate truncation point. We allow the window lag size to be 20 resting upon
the technique of window closing which enables a researcher to make her choice in
the process of learning about the shape of the spectrum instead of relying on any
rules of thumb.17 The estimated spectra of the HP, BK, BN and STSM cycles of real
GDP, consumer real wages and producer real wages, and the cross–spectra between
all GDP cycles and the real wage cycles are shown in Appendix B.
In the following, we focus on the interpretation of the estimated phase angle.
In Figures 4 and 5, the point estimates of the phase lead of the real GDP cycle
over the corresponding consumer and producer real wage cycle, respectively, for all
decomposition methods along with the respective confidence bounds are drawn.18
The frequency range presented here covers all business cycle periodicities, i.e. periods
between 1.5 (frequency of about 1.0) and 8 years (frequency of about 0.2).19 The
relationship between frequency ω and period p is given by the formula: p = 2π/ω. It
16We refer those readers who are not familiar with univariate and multivariate spectral estimation
to, e.g. , Koopmans (1974, Ch. 8) and Priestley (1981, Ch. 6 and 9.5).17The method of window closing is described in Jenkins and Watts (1968, pp. 280–282).18We construct the 90% confidence intervals as described in Koopmans (1974, pp. 285–287).19Following the seminal paper of Burns and Mitchell (1946) this is the commonly used range for
the business cycle length.
20
should be noticed that the vertical axis representing the values of the phase angle is
divided into four regions.20 If the confidence interval covers one of two upper regions
we say that the real GDP cycle significantly leads the real wage cycle. The opposite
holds true if the confidence interval lies in one of the two lower regions. A significant
procyclical behavior of the real wage cycle is indicated by the confidence interval in
the two regions around 0. If, on the other hand, the confidence interval covers the
top or the bottom region we conclude that the real wage behaves countercyclically.
If the confidence interval covers at least three regions, we interpret it as being a “no
information confidence interval”.
In Figure 4, it is apparent that for the consumer real wage the point estimates
of the phase angle display a similar pattern in the case of the HP, BK and BN
cycles. At all frequencies, the estimated phase angle takes on positive values which
suggests a lagging behavior of cycles of the real wage characterized by business
cycle frequencies with respect to the corresponding cycles of real GDP. However,
statistical significance of such a behavior pertains rather to lower business cycle
frequencies. We also observe that for these three decomposition approaches the lower
frequencies (up to about 0.35) are associated with estimates of the phase angle in
the interval [0, π/2]. The significant ones are confined to the frequencies up to 0.25
thereby indicating significant procyclical pattern of the consumer real wage at these
frequencies. In contrast, shorter cycles of the real wage are negatively correlated
with the particular real GDP cycle as shown by the estimated phase angle values
lying above π/2. For the STSM cycles we obtain positive point estimates at lower
frequencies as well. However, we cannot make any statement about the statistical
significance of any estimated phase angle in the whole frequency range. Taking all
findings into account we can conclude that, in general, longer consumer real wage
20The results for each frequency are illustrated on a linear scale which can be obtained through
“straightening” a circular scale connected by the points representing angles π and −π.
21
Figure 4: Phase angle: real GDP and consumer real wage cycles
0.0 0.2 0.4 0.6 0.8 1.0
HP cyclesπ
−π
π/2
−π/2
0
0.0 0.2 0.4 0.6 0.8 1.0
BK cyclesπ
−π
π/2
−π/2
0
0.0 0.2 0.4 0.6 0.8 1.0
BN cyclesπ
−π
π/2
−π/2
0
0.0 0.2 0.4 0.6 0.8 1.0
STSM cyclesπ
−π
π/2
−π/2
0
Notes: The horizontal axis represents (angular) frequency ω.
cycles seem to exhibit a procyclical and lagging behavior, whereas the shorter ones
evolve countercyclically and also react with delay to the actual economic situation.
As for the producer real wage, the estimation results presented in Figure 5 look
almost identical to the ones for the consumer real wage if we consider the estimated
values of the phase angle at lower frequencies. Also, the values in the interval
[−π,−π/2] (HP, BK and BN cycles) at the frequencies above 0.4 (below 4 years)
could serve as an indicator for a countercyclical behavior of the producer real wage.
Despite this similarity to the correlation scheme of the shorter cycles of the consumer
real wage, it can be noted that for shorter cycles the producer real wage seems to lead
22
Figure 5: Phase angle: real GDP and producer real wage cycles
0.0 0.2 0.4 0.6 0.8 1.0
HP cyclesπ
−π
π/2
−π/2
0
0.0 0.2 0.4 0.6 0.8 1.0
BK cyclesπ
−π
π/2
−π/2
0
0.0 0.2 0.4 0.6 0.8 1.0
BN cyclesπ
−π
π/2
−π/2
0
0.0 0.2 0.4 0.6 0.8 1.0
STSM cyclesπ
−π
π/2
−π/2
0
Notes: The horizontal axis represents (angular) frequency ω.
the corresponding real GDP cycle. The main problem, however, is the insignificance
of almost all estimates. Hence, the results for producer real wages in the frequency
domain remain inconclusive.
5 Summary and Conclusions
This paper provides stylized facts about the cyclical behavior of consumer and pro-
ducer real wages in Germany. In order to see whether a robust empirical picture
on real wage behavior emerges, several detrending methods have been applied to
23
both real wage series and real GDP, including the deterministic trend model, the
Beveridge–Nelson decomposition, the Hodrick–Prescott filter, the Baxter–King filter
and the structural time series model. The stochastic properties of the original time
series and the derived cyclical components were analyzed using a set of unit root
tests and other diagnostic tests. Since the cycles generated by the deterministic
trend model violated the stationarity condition, they were excluded from further
analysis.
We then analyzed the comovements of the detrended real wage series with real
GDP in the time domain and in the frequency domain. For both approaches not
only the contemporaneous correlation between real wages and GDP, but also the
lag–lead structure has been taken into account. In the time domain the sample cross–
correlations between the cycle of each of the real wage series and the GDP cycle have
been evaluated. According to our results in the time domain, the contemporaneous
correlation between the real wage and GDP is statistically insignificant, with the
exception of the cycles from the structural time series model. In the latter case we
found a negative contemporaneous correlation. Regarding the lead–lag structure, the
consumer real wage displays a procyclical pattern and lags behind the business cycle.
The strongest reaction to the actual economic situation can be observed between the
5th and the 11th quarter. For the producer real wage all estimated contemporaneous
cross–correlations are statistically insignificant. Furthermore, although there is a
similar cyclical pattern as in the case of the consumer real wage, the sample cross–
correlations at the leads of the real wage are not as high.
In the next step, we analyzed the comovements in the frequency domain. The
great advantage of an analysis in the frequency domain is that it allows to assess
the relative importance of particular frequencies for the behavior of real wages. We
followed the non-parametric approach to the estimation of spectra and cross–spectra.
The analysis of the phase angle for the consumer real wage shows that the observed
24
cyclicality depends on the frequency range under consideration. All decomposition
methods for which we got statistically significant results reveal a similar pattern.
The consumer real wage is lagging the real GDP cycle. For shorter time periods
up to about three years, the consumer real wage shows an anticyclical behavior,
whereas for longer time spans a procyclical behavior can be observed. However, for
the producer real wage the results in the frequency domain remain inconclusive.
Our results for consumer real wages are in line with an economy that is char-
acterized by wage stickiness in the short run. For example, an economic upswing
could first lead to a decline in real wages because of rising prices and rigid nominal
wages. In the longer run, nominal wages are adjusted upwards eventually leading
to a rise in real wages as well.
25
A Data Selection
We use quarterly data for Germany from 1970.Q1 to 2009.Q1 (157 observations).
All series that served to generate the project data, except for working hours, have
already been available as seasonally adjusted data based on the Census-X12-ARIMA
procedure.
Real GDP
In order to obtain the real GDP series we used the price adjusted chain index with
the base year 2000 (source: Deutsche Bundesbank, series JB5000). The raw data
before 1991.Q1 referred to West Germany and after 1991.Q1 to unified Germany.
The index series has already been linked over the annual average for 1991. We
multiplied the index with the nominal GDP in 2000 and divided it by 100 (source
of nominal GDP: Statistisches Bundesamt, GENESIS online database).
Real wages
We obtained the real wage series on the basis of gross wages and salaries (source:
prior to 1991.Q1 Statistisches Bundesamt, Beiheft zur Fachserie 18, Reihe 3; from
1991.Q1 on Statistisches Bundesamt, GENESIS online database). Since we were
interested in hourly real wages, we divided this series by total working hours of the
domestic labor force. The data for working hours from 1970.Q1 to 1991.Q4 referred
to West Germany (source: Statistisches Bundesamt, Erganzung zur Fachserie 13,
Reihe S.12) and from 1991.Q1 on to unified Germany (source: Statistisches Bun-
desamt, GENESIS online database). After seasonal adjustment with the Census-
X12-ARIMA procedure we linked both series over the annual average for 1991. The
nominal hourly wage has been deflated with the consumer price index (CPI) or the
producer price index (PPI) in order to generate the respective real wage series. The
source of both price indices is Deutsche Bundesbank (CPI: series USFB99, PPI:
series USZH99).
26
B Figures: nonparametric spectral estimates
Figure B.1: Spectra of the real GDP cycles
HP
0.0 0.5 1.0 1.5 2.0
0.5
1.0
1.5
HP BK
0.0 0.5 1.0 1.5 2.0
0.5
1.0
1.5 BK
BN
0.0 0.5 1.0 1.5 2.0
0.5
1.0
1.5 BN STSM
0.0 0.5 1.0 1.5 2.0
1
2
3 STSM
Notes: The horizontal axis represents the (angular) frequency ω.
The values on the vertical axis have been multiplied by 104.
27
Figure B.2: Spectra of the consumer real wage cycles
HP
0.0 0.5 1.0 1.5 2.0
0.25
0.50
0.75
1.00 HP BK
0.0 0.5 1.0 1.5 2.0
0.25
0.50
0.75
1.00 BK
BN
0.0 0.5 1.0 1.5 2.0
1
2
BN STSM
0.0 0.5 1.0 1.5 2.0
1
2
3 STSM
Notes: The horizontal axis represents the (angular) frequency ω.
The values on the vertical axis have been multiplied by 104.
28
Figure B.3: Spectra of the producer real wage cycles
HP
0.0 0.5 1.0 1.5 2.0
1
2
3
HP BK
0.0 0.5 1.0 1.5 2.0
1
2
3 BK
BN
0.0 0.5 1.0 1.5 2.0
1
2
3 BN STSM
0.0 0.5 1.0 1.5 2.0
2
4
6STSM
Notes: The horizontal axis represents the (angular) frequency ω.
The values on the vertical axis have been multiplied by 104.
29
Figure B.4: Cospectra and quadrature spectra between the real GDP cycles
and the consumer real wage cycles
HP
0.0 0.5 1.0 1.5 2.0
0.00
0.25
CospectrumHP HP
0.0 0.5 1.0 1.5 2.0
0.00
0.25
0.50
0.75 Quadrature spectrumHP
BK
0.0 0.5 1.0 1.5 2.0
0.00
0.25
0.50BK BK
0.0 0.5 1.0 1.5 2.0
0.00
0.25
0.50
0.75BK
BN
0.0 0.5 1.0 1.5 2.0
0.0
0.5
1.0
1.5BN BN
0.0 0.5 1.0 1.5 2.0
0.00
0.25
0.50 BN
STSM
0.0 0.5 1.0 1.5 2.0
−0.50
−0.25
0.00STSM
STSM
0.0 0.5 1.0 1.5 2.0
0.25
0.75 STSM
Notes: The horizontal axis represents the (angular) frequency ω.
The values on the vertical axis have been multiplied by 104.
30
Figure B.5: Cospectra and quadrature spectra between the real GDP cycles
and the producer real wage cycles
HP
0.0 0.5 1.0 1.5 2.0
0.00
0.25
Cospectrum
HP
HP
0.0 0.5 1.0 1.5 2.0
0.00
0.25
0.50Quadrature spectrum
HP
BK
0.0 0.5 1.0 1.5 2.0
0.00
0.25BK
BK
0.0 0.5 1.0 1.5 2.0
0.00
0.25
0.50 BK
BN
0.0 0.5 1.0 1.5 2.0
0.0
0.5
1.0
1.5BN BN
0.0 0.5 1.0 1.5 2.0
−0.1
0.1 BN
STSM
0.0 0.5 1.0 1.5 2.0
−0.25
0.00
0.25STSM
STSM
0.0 0.5 1.0 1.5 2.0
0.00
0.25
0.50STSM
Notes: The horizontal axis represents the (angular) frequency ω.
The values on the vertical axis have been multiplied by 104.
31
References
Abraham, K. G., and Haltiwanger, J. C. (1995). Real Wages and the Business
Cycle. Journal of Economic Literature, 33 (3), 1215–1264.
Andrews, D. W. K. (1993). Tests for Parameter Instability and Structural Change
With Unknown Change Point. Econometrica, 61 (4), 821–856.
Beveridge, S., and Nelson, C. R. (1981). A New Approach to Decomposition of
Economic Time Series into Permanent and Transitory Components with Par-
ticular Attention to Measurement of the ’Business Cycle’. Journal of Monetary
Economics , 7 , 151–74.
Brandolini, A. (1995). In Search of a Stylized Fact: Do Real Wages Exhibit a
Consistent Pattern of Cyclical Variability? Journal of Economic Surveys ,
9 (2), 103–163.
Burns, A. F., and Mitchell, W. C. (1946). Measuring Business Cycles. New York:
NBER.
Clark, P. K. (1989). Trend Revision in Real Output and Unemployment. Journal
of Econometrics , 40 (1), 15–32.
Cogley, T., and Nason, J. T. (1995). Effects of the Hodrick-Prescott Filter on the
Trend and Difference Stationary Time Series: Implications for Business Cycle
Research. Journal of Economic Dynamics and Control , 19 , 253–278.
Commandeur, J. J. F., and Koopman, S. J. (2007). An Introduction to State Space
Time Series Analysis. New York: Oxford University Press.
Cuddington, J. T., and Winters, L. A. (1987). The Beveridge-Nelson Decomposi-
tion of Economic Time Series: A Quick Computational Method. Journal of
Monetary Economics , 19 , 125–127.
DeJong, D. N., and Dave, C. (2007). Structural Macroeconometrics. Princeton:
Princeton University Press.
Gomez, V. (2001). The Use of Butterworth Filters for Trend and Cycle Estimation
32
in Economic Time Series. Journal of Business and Economic Statistics , 19 (3),
365–373.
Hart, R. A., Malley, J. R., and Woitek, U. (2009). Real Earnings and Business
Cycles: New Evidence. Empirical Economics , 37 , 51–71.
Harvey, A. C. (1989). Forecasting, Structural Time Series Models and the Kalman
Filter. Cambridge: Cambridge University Press.
Harvey, A. C., and Jaeger, A. (1993). Detrending, Stylized Facts and the Business
Cycle. Journal of Applied Econometrics , 8 , 231–247.
Harvey, A. C., and Koopman, J. S. (1992). Diagnostic Checking of Unobserved-
Component Time Series Models. Journal of Business and Economic Statistics ,
10 (4), 377–389.
Harvey, A. C., and Streibel, M. (1998). Tests for Deterministic versus Indeter-
ministic Cycles. Unpublished Paper: Department of Economics, University of
Cambridge.
Harvey, A. C., and Trimbur, T. M. (2003). General Model-Based Filters for Ex-
tracting Cycles and Trends in Economic Time Series. Review of Economics
and Statistics , 85 (2), 244–255.
Hodrick, R. J., and Prescott, E. C. (1980). Postwar U.S. Business Cycles: An Em-
pirical Investigation. Discussion Paper No. 451, Carnegie-Mellon University.
Jenkins, G. M., and Watts, D. G. (1968). Spectral Analysis and its Applications.
San Francisco: Holden-Day.
Keynes, J. M. (1936). The General Theory of Employment, Interest and Money.
London: Macmillan and Co.
Koopman, S. J., Harvey, A. C., Doornik, J. A., and Shephard, N. (2009). STAMP
8.2: Structural Time Series Analyser, Modeller and Predictor. London: Tim-
berlake Consultants.
Koopman, S. J., Shephard, N., and Doornik, J. A. (2008). SsfPack 3.0: Statistical
33
Algorithms for Models in State Space Form. London: Timberlake Consultants.
Koopmans, L. H. (1974). The Spectral Analysis of Time Series. London: Academic
Press.
Messina, J., Strozzi, C., and Turunen, J. (2009). Real Wages over the Business
Cycle: OECD Evidence from the Time and Frequency Domains. Journal of
Economic Dynamics and Control , 33 (6), 1183–1200.
Miller, S. M. (1988). The Beveridge-Nelson Decomposition of Economic Time Series:
Another Economical Computation Method. Journal of Monetary Economics ,
21 , 141–142.
Morley, J. C. (2002). A State-Space Approach to Calculating the Beveridge-Nelson
Decomposition. Economic Letters , 75 , 123–127.
Murray, C. J. (2003). Cyclical Properties of Baxter-King Filtered Time Series.
Review of Economics and Statistics , 85 (2), 472–476.
Newbold, P. (1990). Precise and Efficient Computation of the Beveridge-Nelson
Decomposition of Economic Time Series. Journal of Monetary Economics ,
26 , 453–457.
Priestley, M. B. (1981). Spectral Analysis and Time Series (Z. W. Birnbaum and
E. Lukacs, Eds.). London: Academic Press.
34